Brian Del Rosario, a software engineer and part-time city council member in a small Utah town, uses artificial intelligence chatbots to handle everything from meal planning to managing schedules. In some of these conversations, he revealed that he has a spouse and three children.
Then, after he and his wife separated, Del Rosario had to mention it to a chatbot so that his wife wouldn’t be included when planning future trips. But once he does, the chatbot zeroes in on the divorce.
When he asks for help managing his schedule, it’s a sign that he may be exhausted from the divorce. When he vented about a frustrating day at work, his stress was linked back to the divorce.
He said he told the chatbot, “I’m not asking you to comment on my divorce every chance you get.” Del Rosario said the chatbot “wasn’t going to let it go.”
One of the best things about chatbots is that they have long memories and can learn more about you from one conversation to the next. The result is a smarter assistant that understands your writing style, remembers your dietary restrictions, or picks up a project where you left off.
But this wonderful memory also has some drawbacks. It can be mired in misunderstandings or outdated information. It may feel invasive. This can make it harder to turn over a new leaf in your life.
In other words, you may be over it, but your chatbot isn’t.
eternal fact
Since ChatGPT introduced in-memory in early 2024, Google’s Gemini, Anthropic’s Claude and Microsoft’s Copilot have all added their own versions. Their approaches differ, but the basic idea is the same: the chatbot remembers what you tell it and uses it to shape future responses. Google’s personal intelligence feature can even pull information from a user’s Gmail, photos and YouTube activity.
One problem: It might remember things that have nothing to do with you. For example, a health concern raised on behalf of a child or parent may be mistaken for your own problem. Ask about your child’s ADHD symptoms, and a few weeks later, when you ask for productivity advice, the chatbot might tailor its suggestions based on what attention difficulties it thinks you have.
Alphabet-owned Google admits A version of this question in the blog post Regarding its personal intelligence capabilities, a hypothetical situation was described where the system saw hundreds of photos of a user on a golf course and assumed he enjoyed the sport, when in fact he only did it for his son. (In a hypothetical example, the user would have to tell the chatbot that he doesn’t like golf.)
Google said it launched a feature that allows users to maintain personalization but prevents specific information from reappearing in conversations with chatbots. Meanwhile, OpenAI says it has released an update for Plus and Pro users that improves the way details are looked up and retrieved from memory. Microsoft says users can update or delete specific memories (or all content, if they choose). They can also turn off personalization and memories completely at any time. Human declined to comment.
In shared accounts, which are common between partners or in small businesses, the risks are multiplied. A person uses a chatbot to polish a resume. Later, when someone else in the account asks an unrelated question, the chatbot might reference that person’s recent career moves or suggest next steps for a job search as part of the answer.
Memories also become stale. Let’s say you told the chatbot you were training for a marathon six months ago. You’ve since torn your ACL, but you’ve never spoken about it. Now, every meal plan and fitness advice is calibrated for highly active people. The advice you’re following was made for a version of you that no longer exists.
Really frustrating
Del Rosario experienced something similar. He had mentioned that he was trying to lose weight, and the chatbot started mentioning that fact everywhere, including when he was looking for restaurant recommendations out of town.
“It’s like, ‘Thank you for bringing a buzz to my holiday season,'” he said. “I wasn’t actually planning on sticking to my diet on this trip.”
Likewise, Mike Taylor, a technology consultant at Every, a media and software company in Hoboken, New Jersey, mentioned to the chatbot that he was a British expat. He said the chatbot then recommended a “suitable pint” at a local pub, but he didn’t think the tip was useful. “I’m here for the dive bars in America, not the ones in England,” he said. “That’s why I moved here.”
Taylor has turned off the AI chatbot’s memory so that he knows exactly what’s influencing its response to him. “LL.M. [large language model] It doesn’t know who you are, so it doesn’t bias the results you get,” he said.
In fact, AI memory can shape outcomes in ways that are imperceptible. Joshua Joseph, chief artificial intelligence scientist at Harvard University’s Berkman Klein Center, compared the effect to that of social media feeds, and how the few posts you linger on can quietly reshape everything you see subsequently.
Let’s say you mention in a completely different conversation that you’re stressed about money. A few weeks later, you ask a chatbot for career advice, and it steers you toward a high-paying job instead of a role you might be a better fit for because it “knows” you’re financially anxious. You’ll never know why this suggestion feels uncomfortable, because the chatbot doesn’t mark which memories it’s tapping into.
“It definitely leads, it definitely affects the outcome,” Joseph said. “And we really don’t know how many.” He turned off the memory function on his chatbot account.
Chatbots that remember everything will also make it harder for people to let go of their past. Lucy Osler, a philosophy lecturer at the University of Exeter who studies how artificial intelligence shapes cognition, said chatbots use facts to construct a narrative about who you are and feed that narrative back to you as if it were fact. If you tell your chatbot that you feel insecure and anxious, it will view you that way and may keep reminding you of it even after you’ve moved away.
“It might validate some of the self-narratives I have about myself and make them sound more authentic,” Osler said. “They can lock you up.”
negative pattern
It can be upsetting to be reminded months later that you are feeling anxious. But it can also cause real harm. Chatbots are designed to be enjoyable, building on your version of reality rather than questioning it. Osler said this allows the chatbot to reinforce delusional thinking.
Such concerns prompted the Electronic Privacy Information Center to draft legislation on the safety of chatbots for teenagers, who are particularly susceptible to the ingratiating tendencies of these tools. A key provision requires that memories be wiped between sessions, specifically to prevent the chatbot from establishing harmful mental states over time.
Del Rosario eventually came up with his own method. After his divorce kept seeping into irrelevant conversations, he started using separate chatbots for different parts of his life, using anonymous mode for anything sensitive.
He still values it when it works, like when it knows his kid needs a car seat for a road trip, or when it reminds him he has a lot to do. His mom passed away two years ago, and between divorce, kids, and work, the chatbot is sometimes the only thing that gets the full picture.
“It feels good to be seen, even if it’s being seen by an AI chatbot,” he said.
The main AI assistants let users turn off memory entirely, and each offers some way to view, edit, or delete stored content. But many users don’t understand these new features or never check these settings. Here are some tips to make sure memory functions in your favor:
Find out what your chatbot knows about you. Go to your settings (in ChatGPT it’s under “Personalization”), in Claude and Gemini it’s in the “Memory” section and check what’s there and delete it if necessary.
Temp chat is available for anything sensitive. Or just don’t tell it anything sensitive. The controls and names for casual chats may vary from app to app, but there’s usually a button at the top of the page.
Try turning memory off completely and see how it changes your results. You may find that you prefer this trade-off.
Think of it like a social media profile. It’s worth checking in and updating occasionally because it shapes what you see whether you look at it or not.
write to reports@wsj.com
